AI news · Tuesday, October 6, 2026
OpenAI drops hundreds of AI-generated math proofs on GitHub today
OpenAI is trying to smooth things over with mathematicians after repeatedly bulldozing academic norms. The company just released 722 manuscripts detailing solutions to hundreds of long-standing mathematical problems, including those produced by its latest, unreleased model. Following criticism that its previous releases were just glorified marketing, OpenAI is now publishing these results in a GitHub repository with protocols for citations and machine-verifiable Lean proofs. To manage the backlash, the company is consulting a new advisory group of elite mathematicians, though some insiders describe the process as messy and still fundamentally rushed. Mathematicians remain wary, especially after earlier instances where OpenAI allegedly scooped human researchers and tried to exclude them from credit on breakthroughs. The model reportedly solved these problems using the equivalent compute of three hours of ChatGPT Pro thinking, a pace that has left many in academia feeling like they’re playing catch-up against a runaway bulldozer.
Europe is pushing back with its own frontier-level ambitions. Mistral AI just launched "Le Chonk," a massive one-trillion-parameter model aimed at cybersecurity and chip design. Unlike its American rivals, Mistral is leaning into an open-weight strategy, though it’s keeping the model under wraps for three weeks of safety testing to prevent malicious use. The lab is positioning itself as a secure, "third way" alternative to American-controlled closed models and Chinese alternatives, backed by heavyweights like Samsung and ASML.
Personal AI agents continue to stir up privacy and trust issues as they move from personal inboxes into shared group chats. Startup Instinct is now allowing agents to join group DMs to coordinate logistics, though early users remain nervous about granting these bots "keys to the castle" after a string of hallucinations and unasked-for account access incidents. The privacy-first approach is also gaining traction as a feature: LibreOffice officially decided to exclude AI entirely to ensure user data never leaves the local machine. This stands in stark contrast to Meta’s Muse or Google’s Gemini, which increasingly require users to trade deep, sensitive profile access for agentic convenience. As these agents proliferate, the definition of what constitutes a "recording" is blurring, with companies claiming their hardware "doesn't save" data even while it processes intimate details of your daily life.
The quick hits
- OpenAI released hundreds of AI-generated mathematical proofs on GitHub to address past criticism — an attempt to align with academic standards after months of messy, high-profile conflicts.
- French lab Mistral AI introduced a one-trillion-parameter model called 'Le Chonk' — the latest attempt to build a powerful open-weight competitor that sidesteps the current U.S.-China AI divide.
- Personal AI agents are moving into group chats, but users remain skeptical — high-profile privacy blunders and 'hallucinations' have led many to delete these tools to keep their inboxes secure.
- LibreOffice is betting that 'no AI' is a competitive feature — the nonprofit is keeping generative AI out of its software to guarantee that user data stays strictly on-device.